Symmetric probabilistic alignment for example-based translation

EAMT 2005 Conference Proceedings 153 Symmetric Probabilistic Alignment for Example-Based Translation Jae Dong Kim, Ralf D. Brown, Peter J. Jansen, Jaime G. Carbonell Language Technologies Institute Carnegie Mellon University Pittsburgh, PA 15213 {jdkim,ralf,pjj,jgc}@cs.cmu.edu Abstract. Since subsentential alignment is critically important to the translation quality of an Example-Based Machine Translation (EBMT) system which operates by finding and combining phrase-level matches against the training examples, we recently decided to de- velop a new alignment algorithm for the purpose of improving the EBMT system’s per- formance. Unlike most algorithms in the literature, this new Symmetric Probabilistic Align- ment (SPA) algorithm treats the source and target languages in a symmetric fashion. In this paper, we describe our basic algorithm and some extensions for using context and posi- tional information, compare its alignment accuracy with IBM Model 4, and report on ex- periments in which either IBM Model 4 or SPA alignments are substituted for the aligner currently built into the EBMT system. Both Model 4 and SPA are significantly better than the internal aligner and SPA slightly outperforms Model 4 despite being handicapped by incomplete integration with EBMT. 1. Introduction The example-based translation system (Brown, 2000; Brown et al., 2003) used for the experi- ments described in this paper is internally a multiengine system (Frederking et al., 1994) consisting of a phrasal EBMT engine and word- for-word dictionary lookup which are combined using a language model-driven lattice search equivalent to the decoder used in many statisti- cal machine translation systems1. The phrasal EBMT engine finds all partial matches between input sentences and the bilingual training cor- pus and then uses sub-sentential alignment be- tween the halves of the retrieved examples to determine the corresponding translations. This makes the alignment process a critical factor in the quality of the system’s translations. Incorrect alignments result in missing or ex- traneous words in the translation hypothesized for a particular fragment of the input. This can 1 Other engines can be added to the system, but were not for these experiments, and are generally not available anyway. cause critical information to be omitted, irrele- vant words to be “hallucinated”, or even invert the meaning of the text. Because retrieved train- ing examples with low confidence alignments are not used, lack of a reliable alignment can result in the best translation for a given frag- ment being missed, or even no translation at all for that fragment of input. The latter is in fact a significant issue in the current implementation of EBMT, as it is quite common to have phrasal matches for 90% of the input words yet get translations for only 75% of the input words – the remaining matches did not produce any alignments that were considered satisfactory. Because we recognized that the existing alignment algorithm is a substantial bottleneck in the system’s translation performance, we de- cided to investigate a new algorithm and its ef- fect on translation quality. In the remainder of this paper, we describe the existing aligner, the new SPA method, an …

Content truncated to 500 words